large-traversaal/HYDRA-M3-V0

---
license: mit
task_categories:
- question-answering
- text-generation
language:
- en
tags:
- finance
- multimodal
- multihop
- rag
- 10-K
- financial-analysis
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: "Dataset/finalized_dataset.jsonl"

---

# MMM_HYDRA: Heterogeneous Yielding Dataset for Reasoning Across - Multi-hop, Multimodal, Multicompany

## Dataset Description

**MMM_HYDRA** is a benchmark dataset for evaluating Retrieval-Augmented Generation (RAG) systems on complex financial document analysis. The dataset contains 200 carefully curated questions with answers extracted from 99 unique corporate 10-K filings across 15 industry sectors.

### Key Features

- **Multi-Company**: 54 questions (27%) span multiple corporate entities requiring cross-company analysis
- **Multimodal**: Text, images, and tables from financial documents
- **Multihop**: Requires reasoning across multiple document sections and sources
- **Real-World**: Based on actual SEC 10-K filings from major corporations

### Dataset Statistics

- **Total Questions**: 200
- **Unique Documents**: 99 corporate 10-K filings
- **Industry Sectors**: 15 (Tech Giants, Fast Food, Healthcare, Retail, etc.)
- **Multi-Company Questions**: 54 (27%)
- **Single-Company Questions**: 146 (73%)
- **Average Question Length**: 142 characters
- **Average Answer Length**: 543 characters

### Question Distribution

**By Type:**
- Long Answer: 106 questions (53%)
- Short Answer: 94 questions (47%)

**By Industry Sector:**
- Tech Giants: 43 questions (21.5%)
- Fast Food: 23 questions (11.5%)
- Delivery & Groceries: 22 questions (11%)
- Beverages: 22 questions (11%)
- Entertainment: 20 questions (10%)
- Retail: 19 questions (9.5%)
- Airlines: 12 questions (6%)
- Healthcare: 9 questions (4.5%)
- Others: 30 questions (15%)

**By Solution Modality:**
- Text Only: ~50%
- Text + Table: ~25%
- Text + Image: ~15%
- Table Only/Image Only: ~10%

**By Document Count:**
- 2 documents: 150 questions (75%)
- 3 documents: 40 questions (20%)
- 4 documents: 8 questions (4%)
- 1 document: 1 question (0.5%)

### Top Companies in Dataset

Most frequently referenced companies:
1. Meta (42 references)
2. Alphabet/Google (28 references)
3. Burger King (17 references)
4. Uber (17 references)
5. Wendy's (12 references)
6. DoorDash (12 references)
7. Best Buy (12 references)
8. McDonald's (11 references)

## Dataset Structure

### Data Configurations

This dataset includes multiple configurations accessible through the viewer:

- **default**: Main dataset with questions and answers (`dataset/dataset.json`)
- **images_metadata**: Metadata for associated images (`dataset/images_csv`)
- **categories**: Category information (`dataset/categories.csv`)
- **tables**: Structured table data (`tables/all_tables.json`)

### Data Fields

- `id`: Unique identifier for each question
- `q_no`: Question number
- `Subset`: Industry sector/category
- `Multi company ?`: Boolean indicating if question spans multiple companies
- `Number of Docs`: Number of source documents required
- `Question`: The question to answer
- `Answer`: Ground truth answer
- `Question Type`: Short Answer or Long Answer
- `Solution Requires`: Modality required (Text Only, Text + Table, Text + Image, etc.)
- `Context(s)`: Relevant document sections (text)
- `Context Images`: Associated image references
- `Context tables`: Associated table references
- `Related Images`: Image identifiers
- `Related tables`: Table identifiers
- `Number of Pages`: Page count of source documents
- `Solution in Page(s)`: Specific pages containing the answer
- `Sources DOCS`: List of source document filenames

#
## Benchmark Use Cases

This dataset is designed to evaluate:

1. **Multi-document Reasoning**: Questions require synthesizing information from 1-4 documents
2. **Multimodal Understanding**: Integration of text, tables, and images from financial documents
3. **Cross-company Analysis**: Comparing metrics and strategies across different companies
4. **Financial Domain Knowledge**: Understanding of business terminology and financial concepts
5. **Long-form Generation**: Producing detailed, accurate answers to complex questions

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{mmm_hydra_2025,
title={MMM_HYDRA: Multi-Company Multimodal Multihop Financial Reasoning Benchmark},
year={2025},
publisher={Hugging Face},
howpublished={\url{https://huggingface.co/datasets/your-username/MMM_HYDRA}}
}
```

## License

This dataset is released under the MIT License.

## Acknowledgments

This dataset is built from publicly available SEC 10-K filings and is intended for research and evaluation purposes.
This dataset is released under the MIT License.